Headline

Product | Agile | Startups

About

Product | Generative AI | Agentic Systems | 0→1 Product Builder | PM’s Awardee | 7 Patents Over the last 11 years, my focus has been on building the technical infrastructure behind large-scale enterprise platforms and generative AI applications. Most of my day-to-day work lives at the intersection of systems architecture, data retrieval, and performance economics. Using Agile methodologies to coordinate cross-functional engineering teams, I ran the end-to-end product lifecycle for Domino, a multimodal Gen-AI video platform, scaling it to a syndicated reach of over 1 billion users. The core challenge there was context engineering - essentially configuring Pinecone-backed RAG pipelines with custom semantic chunking to parse data from messy source URLs before any video rendering ever started. I tend to view foundation models and AI features as distributed infrastructure components rather than standalone tools; if you want them to survive real-world scale, you have to prioritize system judgment and latency optimization. When handling traffic across ecosystems seeing 450M+ monthly sessions, like we did at Session AI, you cant just pass unstructured inputs to a model and hope for the best. I focused heavily on designing the agentic workflow architecture, mapping out strict tool-calling boundaries and setting up deterministic fallback architecture. If an LLM times out or returns garbage data, the core decision engine needs hard-coded translation logic to step in instantly and keep user-facing latency below 100ms.That mindset comes from a deeper background in platform modernization and cloud platforms. During my time at Wayfair, I led Agile sprints to transition legacy batch systems into event-driven microservices on modern cloud infrastructure, breaking down monolithic data footprints into clean service boundaries that ultimately unlocked $127M in supply chain demand. Because AI behavior is inherently probabilistic, a huge chunk of my work now relies on building automated evaluation systems and tracking real-time observability metrics to catch regressions or prompt drift before code deploys. I spend a lot of time on technical stakeholder translation, working closely with dev and data science teams to balance model token costs against real business outcomes. Beyond shipping code, I also hold 7 patents for technical innovations and received the Prime Ministers Award for engineering and innovation in 2023.

Location

Pune District

AI stack

aistudio

claude code

codex

geminicli

hugging face

langsmith

Work experience

Senior Product Manager at Session AI, Jan 2025 - Present · 1 yr 10 mos, Mumbai, Session AI spots purchase intent from on-site behaviour alone, no personal data, no third-party cookies. I ran the AI agent platform across three enterprise retail deployments, including a Fortune 1 retailer, client properties that together see 450M+ site visits a month. It scores every visitor, anonymous ones included, within 5-clicks, and acts in under 100ms. Led AI agent design and testing with a 15-person team across engineering, data science and design, and owned the harness around our LLM agent: which tools it could call, what it was allowed to publish, and how we checked its work before anything went live. Kept real-time decisioning under 100ms with defined SLOs at Fortune 1 traffic levels. Partnered with Data Science on the fourth and fifth-click propensity models (LSTM), hardened the signal pipeline and policy rules for reliability, and tuned thresholds myself when our largest account escalated latency. Designed the agent so the LLM never sat in the request path. The agent (Bedrock) works at design time: a marketer describes a campaign, it proposes segments, offers and an experiment plan, and the harness compiles that into a versioned action policy that the v1 predictive offer engine (SageMaker, drift monitoring from day one) runs deterministically; No model call ever showed up in a p99 trace, which is what let us put it in front of a Fortune 1 account (Walmart). Cut campaign setup from 5 days to 30 minutes, a 10x return on the platform build, by making agent flows self-serve for the retailers' own marketing teams: event signals, feature schemas, embeddings, segmentation logic, action policies. Gave the agent a small, boring toolset (segment stats, policy simulation, offer drafting, experiment creation, publish) and a per-retailer context pack with brand rules and offer limits, so marketers weren't re-explaining the basics every campaign. Setup that used to need our services team now doesn't.

Lead Product Manager at Rizzle, Apr 2022 - Jan 2025 · 2 yrs 10 mos, Pune, Maharashtra, India, Owned end-to-end lifecycle of AI-first creative tools for B2B content creators, from discovery to launch, leading a 20+ cross-functional team (engineering, design, data science). Domino – Text-to-Video Platform Turned articles, scripts, or bullet points into safe, branded videos ready for syndication to 1B+ audiences. Built RAG pipeline (LangGraph/LangChain + Pinecone) with semantic chunking and metadata filters to ground LLM outputs in licensed Getty/partner assets—reducing hallucinations, improving retrieval precision@k. Integrated ElevenLabs TTS, Soundstripe music, and AI-driven scene matching (CLIP embeddings, object detection) to fully automate video assembly. Cut creation time from 24–36 hrs to minutes; reduced costs up to 90%; improved factual accuracy and brand fit. Delivered 1,000+ templates, brand/style kits, and policy guardrails to balance creativity with compliance. Smart Video Workflows Added AI-generated motion graphics, dynamic templates, and B-roll matching, increasing creative throughput 4.5×. Implemented A/B testing loops with CTR/watch-time feedback into retriever for iterative quality gains. CRM & Campaign Integrations Shipped Salesforce sync for lead/customer attributes into campaign targeting, lifting CRM-linked conversions by 15%. SmartEnhance – Image Enhancement SDK B2B iOS/Android SDK + REST APIs to auto-detect subject, sentiment, and aesthetics, then apply ranked LUTs/effects with brand theming. Delivered Core ML on-device scoring, cloud backups, and A/B tests via feature flags and Mixpanel. Raised share rates 10% and daily creations 12%; cut release cycles by 40%. Stack - LangGraph, LangChain, Pinecone, Getty Images, ElevenLabs, Soundstripe, Core ML, AWS, Mixpanel, REST APIs

Product Consultant at Eleos Health, Apr 2021 - Feb 2022 · 10 mo, California, United States, Eleos is an AI scribe for therapy. It records or ingests a session, drafts a clinical note (SOAP/DAP), the clinician reviews/edits, and the note is pushed into the clinic’s EHR. What I owned - The whole flow - record to draft to review to EHR. I set the rules for what a “good draft” must contain (Interventions, Risk, MSE, Plan), defined the interfaces between services, wrote PRDs/release notes, and kept on-call playbooks so pilots ran smoothly. Defined “good draft” standards (Interventions, Risk, MSE, Plan) and the service contracts around them; wrote PRDs and on-call runbooks so pilots stayed stable. Redesigned the review UX around what therapists actually edit and pulled prior goals/plans into context to ~31% fewer edits per 1k tokens and median time-to-sign at 6.7 min (42 clinicians, ~7.9k notes, 12 weeks). Handled Quality trade-offs - early dips on crisis notes to added dry-run preview + clinician override; restored quality without slowing the happy path. HIPAA guardrails - PHI-redaction quality checks, admin toggles (delete-after-scribe / never store audio), encryption in transit/at rest, and full audit logs. Why it mattered to clinicians - Notes got faster to finish. We redesigned the review screen around the sections therapists actually tweak and pulled in the last note/goals so drafts weren’t generic. Impact: edits/1k tokens –31% and median time-to-sign –6.7 minutes (42 clinicians, ~7.9k notes, 12 weeks). Why it mattered to clinics - The EHR step became predictable. We added a dry-run preview, idempotent writes, and a safe fallback to structured copy so no one got stuck. Reliability (P95) - autofill 97.4% success, 2.1% rollbacks, 2.9s fill time. Stack - Python/FastAPI, Postgres, S3 (KMS); React/TypeScript review app; ASR via Deepgram & Whisper; LLM drafting (GPT-4 primary, Claude fallback); light retrieval of prior goals/plans (pgvector/Pinecone).

Senior Product Manager at Wayfair, Mar 2019 - Apr 2022 · 3 yrs 2 mos, Boston, MA, Led the full release of CastleGate Products, transitioning from MVP to full version. Introduced key features like Factory selection, MOQ, and integrated with RCS, resulting in a 20% increase in process efficiency and improved inventory accuracy, contributing to an estimated $120M in forecasted demand. Directed the successful rollout of the Replenishment Catalog System (RCS), replacing the legacy RMS. Executed a phased rollout plan, enhanced the catalog, unlocking 1.7% of forecasted target item demand ($127M), and implemented automated updates, saving 80+ person-hours/month. Enhanced supplier contract flexibility by enabling simultaneous fixed and floating ocean contracts. Optimized order creation based on Expected Cargo Ready Date (ECRD) and dynamically populated origin ports, improving accuracy and user experience.

Product Manager at Fitterfly, Feb 2016 - Mar 2019 · 3 yrs 2 mos, India, Developed and launched the Mental Health Journey Tracker as part of Fitterfly's Digital Health Suite - Mood/somatic check-ins, behavioural activation, exposure ladders, flare plans. After cutting ~40% low-value screens and adding light gamification + rules-based personalisation (with a simple bandit for module order), session completion rose 25% and D30 retention moved from 18% to 29%. Streamlined the coach workflow with intent classification (logistic regression) and rule-based nudges; crisis keywords always routed to humans. Result - ~40% faster time-to-first-response and ~4% of members proactively surfaced for clinical review. Embedded PHQ-9 / GAD-7 / WHO-5 and adherence analytics; shipped lean member/provider/employer views showing utilisation and outcome deltas that actually moved renewals. Led end-to-end API integrations for LSQ CRM Platform, managing multiple CRM integrations and growth initiatives, resulting in a 40% month-over-month increase in sales engagements. Introduced and integrated Payments and BNPL solutions, driving a consistent 33% spike in average monthly sales.

Education

Indian Institute of Technology, Bombay

Certificate in Product Management and Strategy, Product Management, The Wharton School